Twenty languages in two days. Zero lines written by hand.

A YouTube video with over 100,000 views should be an asset. But six months after publication, Johan's 30-minute guide to a popular Obsidian plugin had become a problem. During the months that had passed, the software had been updated 41 times. The search syntax had changed, the settings had been reorganized twice. Features for calendar, tasks, and manual sorting did not even exist when the original was recorded. New viewers were met with a walkthrough that was no longer accurate.

That is why Johan recorded the sequel: "Mastering Notebook Navigator 3". The result was 64 minutes divided into 14 chapters. The starting point for the dubbing was crystal clear. The movie was published, the English speech was transcribed into 580 subtitle blocks, and machine-translated subtitles were live in 20 languages.

The audio, however, was missing. And the list of requirements was clear. Every language needed a complete audio track, exactly as long as the movie. The speech had to follow the subtitles — because the subtitles showed what was happening on screen at that moment. The voice had to be Johan's own, trained in ElevenLabs. Most important of all was control: we needed to be able to pick out and redo a single line without being forced to render a whole hour of audio again.

So we stepped out of the language and into the structure. We had a script read the 580 English blocks and melt them down into 106 whole sentences. Each sentence was tagged with the right chapter and a blocklist for words that could not be touched. A central glossary kept order: Obsidian's "vault" is a workspace in the app, not a bank vault. Menus and buttons like "Settings", "Restricted mode", and "Browse" should be called exactly what was on screen, regardless of language.

Then we spun up one translator agent per language. The instructions were precise, down to pronouns: Brazilian Portuguese had to use "você", European Portuguese "tu", and Ukrainian the formal "ви". The agents translated sentence by sentence in full context, and then cut the text back across the original block boundaries.

Right after, a reviewer agent stepped in. It compared the new, split subtitles with the English source, line by line. A script monitored the timecodes and made sure the text length was right: 50 characters for a Latin line, 32 for Japanese, and only 24 for Chinese.

When the dust settled, the reviewers had logged 2,531 corrections. Seven out of ten were about segmentation — moving the actual break point so the viewer gets to pause their breath where the language demands it. French and German got the most changes because they were used as testbeds. Farsi got its very own extra round in full context to straighten out the mess from the start.

Only when a total of 11,600 blocks passed the tests without a single warning, did we push the button for the audio.

The audio came from ElevenLabs' multilingual model, fed with Johan's cloned voice. Around it, we built a pipeline in Python that sent one API call per line and then cached the audio file based on a unique hash of the text and settings.

That cache was crucial. If the credit balance ran out in the middle of a generation, the system froze, ready to pick up the thread again as soon as the card was charged. And if we later discovered that a single sentence sounded wrong, we could correct it and just ask for that specific audio clip again, along with its neighbors so the tone matched at the seams. Without that detailed control, the project would never have been completed in two days.

Finally, a renderer took over. It stacked the lines on the movie's timeline, measured and compared against the publication. Out came a crash report for each track: exact length, deviations against the subtitle start times, warnings about clips that risked clipping, and pure audio levels. We did not compromise; every language went through the exact same check.

The error was obvious: we put the time gap in the wrong place. The extra time should not be stretched over the vowels, it should be in the pauses.

We fired up Anthropic Fable 5 in Claude Code to rewrite the renderer. In the new version, the speech was locked exactly at the time the subtitle required. The speed was allowed to slow down by a maximum of five percent, which the ear barely perceives. The rest of the time was squeezed in where we humans actually rest — at periods, commas, and colons. And since ElevenLabs delivers character-level timestamps, we could place the pauses exactly right.

If there was still time left over, it was left as pure silence, usually where the original movie also took a pause to show something on screen. So the silence would not turn off the mood in the room, a faint pink noise was embedded as room tone.

Thanks to the caching system, we could throw the entire re-rendering of the 20 languages at existing audio files. The update did not require a single new API call to the voice model.

The suspicion was raised immediately: if German had this problem, what was hiding in the other languages? Since none of us speak Thai, Ukrainian, or Farsi, we turned to the agents again. Four auditors got to plow through all twenty spoken scripts in search of mispronunciations, linguistic collisions, and errors that had been inherited in the translations.

The net was tight. The agents found that "Ctrl" was spelled out pretty much everywhere. The number "15,000" was pronounced as "fifteen point zero" in languages where the comma stands for decimals. The slash in "/Journal" was forgotten. Thai was missing spaces in 66 lines where Latin script broke off. In Turkish, ordinal numbers had been cut right in half by the block boundaries. And four different languages had stiffly translated the word "funny" into "comical", instead of "fun".

The corrections were made in a single batch: 966 lines were regenerated and the timelines were rendered again. Every single respelling resulted in a separate little audio clip for our manual review.

In the end, the project proved one thing. Agentic workflows deliver in production. And they do it even when you are dealing with audio, exact timelines, and twenty languages that nobody on the team understands.

The method worked because it is built on machine-monitored acceptance criteria for length, sync, pronunciation, and terminology. The audio is cached systematically so that a change only costs API money for the exact lines being rewritten. The AI agents do the heavy lifting and audit the volume, but a human always makes the final decisions.

The machinery is driven by standard components: subtitle files, an API for voice generation, tools like ffmpeg, attention to detail, and AI models that wrote the code. That is the kind of workflow we at TokenTek build and automate.